def test_sample(self): q = 0.92 N = 10000 stm = STM(0, 0, 1, 1) stm.biases = [log(q / (1. - q))] x = mean(stm.sample(empty([0, N]))) - q p = 2. - 2. * norm.cdf(abs(x), scale=sqrt(q * (1. - q) / N)) # should fail in about 1/1000 tests, but not more self.assertGreater(p, 0.0001)
def test_pickle(self):
stm0 = STM(5, 10, 4, 21)
tmp_file = mkstemp()[1]
# store model
with open(tmp_file, 'w') as handle:
dump({'stm': stm0}, handle)
# load model
with open(tmp_file) as handle:
stm1 = load(handle)['stm']
# make sure parameters haven't changed
self.assertEqual(stm0.dim_in, stm1.dim_in)
self.assertEqual(stm0.dim_in_nonlinear, stm1.dim_in_nonlinear)
self.assertEqual(stm0.dim_in_linear, stm1.dim_in_linear)
self.assertEqual(stm0.num_components, stm1.num_components)
self.assertEqual(stm0.num_features, stm1.num_features)
self.assertLess(max(abs(stm0.biases - stm1.biases)), 1e-20)
self.assertLess(max(abs(stm0.weights - stm1.weights)), 1e-20)
self.assertLess(max(abs(stm0.features - stm1.features)), 1e-20)
self.assertLess(max(abs(stm0.predictors - stm1.predictors)), 1e-20)
self.assertLess(max(abs(stm0.linear_predictor - stm1.linear_predictor)), 1e-20)
def run_STM(X,y,num_components=3,num_features=20):
if X.shape[0]>X.shape[1]:
X = X.T
if y.ndim==1:
y = y[:,np.newaxis]
if y.shape[0]>y.shape[1]:
y = y.T
model = STM(X.shape[0], 0, num_components, num_features, LogisticFunction, Bernoulli)
model.train(X,y, parameters={
'verbosity':1,
'threshold':1e-7
}
)
yhat = model.predict(X).ravel()
return yhat, model
def test_basics(self): dim_in_nonlinear = 10 dim_in_linear = 8 num_components = 7 num_features = 50 num_samples = 100 # create model stm = STM(dim_in_nonlinear, dim_in_linear, num_components, num_features) # generate output input_nonlinear = randint(2, size=[dim_in_nonlinear, num_samples]) input_linear = randint(2, size=[dim_in_linear, num_samples]) input = vstack([input_nonlinear, input_linear]) output = stm.sample(input) loglik = stm.loglikelihood(input, output) # check hyperparameters self.assertEqual(stm.dim_in, dim_in_linear + dim_in_nonlinear) self.assertEqual(stm.dim_in_linear, dim_in_linear) self.assertEqual(stm.dim_in_nonlinear, dim_in_nonlinear) self.assertEqual(stm.num_components, num_components) self.assertEqual(stm.num_features, num_features) # check parameters self.assertEqual(stm.biases.shape[0], num_components) self.assertEqual(stm.biases.shape[1], 1) self.assertEqual(stm.weights.shape[0], num_components) self.assertEqual(stm.weights.shape[1], num_features) self.assertEqual(stm.features.shape[0], dim_in_nonlinear) self.assertEqual(stm.features.shape[1], num_features) self.assertEqual(stm.predictors.shape[0], num_components) self.assertEqual(stm.predictors.shape[1], dim_in_nonlinear) self.assertEqual(stm.linear_predictor.shape[0], dim_in_linear) self.assertEqual(stm.linear_predictor.shape[1], 1) # check dimensionality of output self.assertEqual(output.shape[0], 1) self.assertEqual(output.shape[1], num_samples) self.assertEqual(loglik.shape[0], 1) self.assertEqual(loglik.shape[1], num_samples)
def test_poisson(self): stm = STM(5, 5, 3, 10, ExponentialFunction, Poisson) # choose random parameters stm._set_parameters(randn(*stm._parameters().shape) / 100.) err = stm._check_gradient( randn(stm.dim_in, 1000), randint(2, size=[stm.dim_out, 1000]), 1e-5) self.assertLess(err, 1e-8)
def test_train(self):
stm = STM(8, 4, 4, 10)
parameters = stm._parameters()
stm.train(
randint(2, size=[stm.dim_in, 2000]),
randint(2, size=[stm.dim_out, 2000]),
parameters={
'verbosity': 0,
'max_iter': 0,
})
# parameters should not have changed
self.assertLess(max(abs(stm._parameters() - parameters)), 1e-20)
def callback(i, stm):
callback.counter += 1
return
callback.counter = 0
max_iter = 10
stm.train(
randint(2, size=[stm.dim_in, 10000]),
randint(2, size=[stm.dim_out, 10000]),
parameters={
'verbosity': 0,
'max_iter': max_iter,
'threshold': 0.,
'batch_size': 1999,
'callback': callback,
'cb_iter': 2,
})
self.assertEqual(callback.counter, max_iter / 2)
# test zero-dimensional nonlinear inputs
stm = STM(0, 5, 5)
glm = GLM(stm.dim_in_linear, LogisticFunction, Bernoulli)
glm.weights = randn(*glm.weights.shape)
input = randn(stm.dim_in_linear, 10000)
output = glm.sample(input)
stm.train(input, output, parameters={'max_iter': 20})
# STM should be able to learn GLM behavior
self.assertAlmostEqual(glm.evaluate(input, output), stm.evaluate(input, output), 1)
# test zero-dimensional inputs
stm = STM(0, 0, 10)
input = empty([0, 10000])
output = rand(1, 10000) < 0.35
stm.train(input, output)
self.assertLess(abs(mean(stm.sample(input)) - mean(output)), 0.1)
def test_glm_data_gradient(self): models = [] models.append( STM( dim_in_nonlinear=5, dim_in_linear=0, num_components=3, num_features=2, nonlinearity=LogisticFunction, distribution=Bernoulli)) models.append( STM( dim_in_nonlinear=5, dim_in_linear=0, num_components=3, num_features=2, nonlinearity=ExponentialFunction, distribution=Poisson)) models.append( STM( dim_in_nonlinear=2, dim_in_linear=3, num_components=3, num_features=2, nonlinearity=LogisticFunction, distribution=Bernoulli)) models.append( STM( dim_in_nonlinear=2, dim_in_linear=3, num_components=4, num_features=0, nonlinearity=LogisticFunction, distribution=Bernoulli)) models.append( STM( dim_in_nonlinear=0, dim_in_linear=3, num_components=2, num_features=0, nonlinearity=LogisticFunction, distribution=Bernoulli)) for stm in models: stm.sharpness = .5 + rand() x = randn(stm.dim_in, 100) y = stm.sample(x) dx, _, ll = stm._data_gradient(x, y) h = 1e-7 # compute numerical gradient dx_ = zeros_like(dx) for i in range(stm.dim_in): x_p = x.copy() x_m = x.copy() x_p[i] += h x_m[i] -= h dx_[i] = ( stm.loglikelihood(x_p, y) - stm.loglikelihood(x_m, y)) / (2. * h) self.assertLess(max(abs(ll - stm.loglikelihood(x, y))), 1e-8) self.assertLess(max(abs(dx_ - dx)), 1e-7)
def test_gradient(self):
stm = STM(5, 2, 10)
stm.sharpness = 1.5
# choose random parameters
stm._set_parameters(randn(*stm._parameters().shape) / 100.)
err = stm._check_gradient(
randn(stm.dim_in, 1000),
randint(2, size=[stm.dim_out, 1000]),
1e-5,
parameters={'train_sharpness': True})
self.assertLess(err, 1e-8)
# test with regularization turned off
for param in ['biases', 'weights', 'features', 'pred', 'linear_predictor', 'sharpness']:
err = stm._check_gradient(
randn(stm.dim_in, 1000),
randint(2, size=[stm.dim_out, 1000]),
1e-6,
parameters={
'train_biases': param == 'biases',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_predictors': param == 'pred',
'train_linear_predictor': param == 'linear_predictor',
'train_sharpness': param == 'sharpness',
})
self.assertLess(err, 1e-7)
# test with regularization turned on
for norm in ['L1', 'L2']:
for param in ['priors', 'weights', 'features', 'pred', 'input_bias', 'output_bias']:
err = stm._check_gradient(
randint(2, size=[stm.dim_in, 1000]),
randint(2, size=[stm.dim_out, 1000]),
1e-7,
parameters={
'train_prior': param == 'priors',
'train_weights': param == 'weights',
'train_features': param == 'features',
'train_predictors': param == 'pred',
'train_input_bias': param == 'input_bias',
'train_output_bias': param == 'output_bias',
'regularize_biases': {'strength': 0.6, 'norm': norm},
'regularize_features': {'strength': 0.6, 'norm': norm},
'regularize_predictors': {'strength': 0.6, 'norm': norm},
'regularize_weights': {'strength': 0.6, 'norm': norm},
})
self.assertLess(err, 1e-6)
self.assertFalse(any(isnan(
stm._parameter_gradient(
randint(2, size=[stm.dim_in, 1000]),
randint(2, size=[stm.dim_out, 1000]),
stm._parameters()))))
